We reviewed the literature on the performance of the blood culture as a diagnostic test and analyzed the data with Bayes' theorem to find the optimal number of cultures to draw. The blood culture is unusually dependent on physician behavior (use of sterile technique, the number and timing of cultures, volume of blood drawn) and on their clinical judgment (estimating the pretest probability of bacteremia, anticipating the causative pathogen, interpreting the results). Because there is no independent "gold-standard" procedure against which to evaluate this test, sensitivity and specificity can only be approximated. Sensitivity can be maximized by doing multiple cultures containing at least 10 mL of blood per set. Specificity can be maximized by adhering strictly to aseptic techniques and by requiring that multiple sets be positive for the series to be considered positive when the anticipated pathogens are also common contaminants. Two or three blood culture sets almost always suffice to establish or rule out bacteremia, although on some occasions obtaining more than three sets of cultures is indicated. One set is rarely, if ever, sufficient.